Voice conversion (VC) models can be used to anonymize speech data while preserving naturalness. Existing models, however, are mainly optimized for English and other high-resource languages, and they fail to generalize well to Arabic due to its distinct phonology. In this work, we comprehensively evaluate existing zero-shot VC systems for Arabic speech, identifying limitations in pronunciation preservation. We then train an Arabic-centric model based on EZ-VC on diverse Arabic speech, and demonstrate strong performance in intelligibility, naturalness, and pronunciation preservation, while maintaining high speaker dissimilarity. Through subjective evaluation, we further demonstrate its robustness to code-switching and cross-dialectal conversion, as well as its generalization to Arabic dialects with limited representation in the training data. Our findings highlight the limitations of zero-shot VC and the importance of target-language adaptation for robust voice anonymization.
Figures & tables
Train
Source Dev
Source Test
Target
H
S
H
S
H
S
H
S
ArVoice
3.9
2
1.4
1
2.2
2
31.0
6
MASC
485.0
0
6.6
101
6.0
90
0.0
0
Casablanca
0.0
0
1.5
-
1.5
-
0.0
0
ArzEn
5.6
14
2.1
11
2.5
13
0.4
2
Mixat
3.2
6
0.2
4
0.2
4
5.9
1
Table 1: Statistics of the labeled corpora used for training, source-speaker development and test, and target-speaker partitions, reporting duration in hours (H) and speaker count (S).
Model
Intelligibility
Privacy
Naturalness
Pronunciation Distance
WER ↓
CER ↓
SIM src↓
SIM tgt↑
EER ↑
UTMOS ↑
All ↓
E ↓
U ↓
P ↓
V ↓
Vec2Wav 2.0
36.0
21.5
0.21
0.52
46.28
2.95
0.24
0.58
0.55
0.24
0.22
YourTTS
35.2
20.5
0.22
0.29
35.88
2.03
0.29
0.58
0.39
0.33
0.26
Diff-HierVC
23.2
12.0
0.29
0.33
42.65
2.54
0.24
0.44
0.37
0.23
0.21
Seed-VC
17.7
7.7
0.22
0.57
49.46
2.71
0.15
0.44
0.34
0.20
0.16
kNN-VC (baseline)
21.1
10.2
0.32
0.55
45.95
2.75
0.14
0.34
0.22
0.14
0.12
Table 2: Results for objective evaluation. Intelligibility is reported using WER and CER (%). SIM src and SIM tgt denote speaker similarity of the converted speech to the source and target speaker, respectively. ASV EER (%) is the equal error rate of an automatic speaker verification system on the converted speech. Pronunciation distance is measured as JSD between phonetic posteriorgrams, across emphatics (E), uvulars (U), pharyngeals (P), vowels (V), and all phonemes.
Models
I ↑
S ↑
N ↑
A ↑
P ↑
Q ↑
Vec2Wav 2.0
4.38
4.15
3.95
4.08
4.52
4.12
Seed-VC
4.70
4.18
4.09
4.27
4.64
4.35
kNN-VC (baseline)
4.81
3.96
4.39
4.62
4.69
4.51
kNN-VC (AR+EN+FR)
4.82
3.93
4.43
4.65
4.63
4.49
EZ-VC (baseline)
4.34
4.19
3.68
3.81
4.27
4.04
EZ-VC (AR+EN+FR)
4.84
4.02
4.53
4.67
4.77
4.55
Table 3: Results for Human Study 1 showing MOS for intelligibility (I), speaker dissimilarity (S), naturalness (N), accentedness (A), style preservation (P), and audio quality (Q).
Models
I ↑
S ↑
N ↑
A ↑
P ↑
Q ↑
(a) Source Monolingual vs. Code-switching (CS)
Egyptian mono
4.95
3.75
4.50
4.80
4.80
4.69
Egyptian CS
4.87
3.84
4.52
4.70
4.67
4.61
Tunisian mono
4.86
3.53
4.42
4.66
4.72
4.51
Tunisian CS
4.82
3.76
4.47
4.51
4.65
4.56
Emirati mono
4.93
4.86
4.59
4.75
4.69
4.66
Table 4: Results for Human Study 2 showing MOS for intelligibility (I), speaker dissimilarity (S), naturalness (N), accentedness (A), style preservation (P), and audio quality (Q). DA denotes dialectal Arabic, and DA x → DA x /DA x → DA y indicate same/different source → target dialects, respectively.
Streaming zero-shot voice conversion struggles to disentangle timbre from linguistic content without degrading utility or inflating latency. Current methods rely on information bottleneck (IB) or speaker perturbation. While IB filters out timbre, it discards prosody, forcing models to explicitly inject features like fundamental frequency. This often requires buffering future frames, creating algorithmic lookahead latency. On the other hand, existing perturbation methods largely overlook the crucial trade-off between timbre leakage and utility preservation. Recognizing this neglected trade-off, we find that the inherent objective of Speaker Anonymization (SA) aligns well with balancing these factors. Thus, we introduce SA as a novel perturbation mechanism to explicitly mitigate timbre leakage while retaining prosodic utility. Crucially, SA's robust representations significantly alleviate the generator's reliance on future context, enabling our strictly causal, zero-lookahead network. Audio samples are available at https://amphionteam.github.io/Zero-VC-demo/.
Yudong Li, Zihao Fang, Junwen Qiu +4
The Chinese University of Hong Kong, Shenzhen · Shenzhen Transsion Holdings Co., Ltd. · Shenzhen Loop Area Institute +1
We present a voice conversion (VC) framework that utilizes K-Nearest Neighbors (KNN) retrieval over WavLM representations to align non-parallel source and target speech, constructing synthetic training pairs for supervised learning. The retrieved segments serve as synthetic inputs, while real target audio provides ground-truth outputs, forming a synthetic-to-real training paradigm that naturally supports multilingual data without requiring parallel corpora or explicit alignment. To ensure consistent target-speaker identity, we incorporate a speaker loss derived from a pretrained speaker verification model. Experiments across multiple languages demonstrate that the proposed approach achieves high naturalness and strong speaker similarity, outperforming competitive VC baselines, despite being trained exclusively on English data. Samples can be accessed at: https://palindromic-vc.github.io.
Detecting synthetic and voice-converted speech remains difficult for low-resource languages with dialectal diversity, where systems must generalize across regional dialects and unseen acoustic channels. We present the UZH-CL submission to the ArA-DF 2026 Shared Task on Arabic speech deepfake detection, covering Track1 (dialect generalization) and Track2 (acoustic robustness). We freeze a W2V-BERT-2.0 backbone and adapt it with Wavelet Prompt Tuning, updating under 1% of parameters, and aggregate multi-layer representations with cross-layer attention and a general attentive-statistics pooling head rather than a specialized graph backend. Complementary detectors are obtained by varying adaptation strategy, training data, augmentation, and encoder family. We find that the two shift types require different fusion regimes: a broad multi-window ensemble for dialect generalization, and a compact, channel-matched, center-crop ensemble for acoustic robustness. Official evaluation yields 1.96% EER on Track1 (6th place) and 1.04% EER on Track2 (3rd place), corresponding to 87% and 96% relative reductions over the XLS-R+AASIST baselines.
Aref Farhadipour, Teodora Vukovic, Petr Motlicek
Department of Computational Linguistics, University of Zurich, Switzerland · Idiap Research Institute, Switzerland · Brno University of Technology, Czech Republic